"Kupol 25" – pneumatics, a neural network, and a plywood pallet versus FPV drones
The main threat to an infantryman in a trench today isn't a sniper or a mortar, but a plastic box with four screws for two hundred dollars. For every such box, a classic Defense you won't have enough: Rocket anti-aircraft complex costs more than a city block of FPV-dronesHence the race for a compact, affordable, and highly automated last-ditch weapon. In May 2026, Russia unveiled one answer to this question: an autonomous turret weighing twenty kilograms, equipped with a neural network, four cameras, and the ability to fire without human intervention.
Phalanx, AK-630, and the logic of the last frontier
Naval CIWS (Close-In Weapon System) emerged at the intersection of two developments. On the one hand, the P-15 family of anti-ship missiles and their Western counterparts demonstrated that classic anti-aircraft artillery A human-aimed system doesn't work well against them. On the other hand, computers capable of calculating lead in real time have matured. The Phalanx Mk 15 entered service in 1980, and the Dutch Goalkeeper in the early 1980s. The Soviet response was parallel: the 30mm AK-630 was installed on ships in the second half of the 1970s, and in the late 1980s, the Kortik appeared, an anti-aircraft missile and artillery system with two 30mm AO-18K automatic guns and eight SAMs on a single turret.
Mark 15 Phalanx Block 1B anti-aircraft artillery system
The engineering essence is the same for all of them. The machine makes the decision, the "detection-aiming-firing" sequence takes place in seconds, and the area of responsibility is a narrow strip of land in the last kilometers before the protected facility. Anything that isn't shot down is left to the repair crew.
The "Dome 25" operates on the same principle, only condensed to the tactical level. For a ship, the "last line" is four to five kilometers. For a trench against an FPV, it's twenty-five meters. The distances differ by two orders of magnitude, but the principle is virtually the same: a rotating base, an automatic feed, a closed-loop firing system without a human in the loop.
What NeuroPVO has assembled: a breakdown of the layout
The NeuroPVO laboratory is a civilian engineering team developing the Kupol 25 system using its own funds and openly publishing its schematics. The photographs with markings (there's also a video from the May 12, 2026, test, and a short clip of the turret targeting a flying quadcopter at a test site) show the entire configuration: a turret on a circular rotating base, two servomotors per axes, four 360° cameras along the perimeter, a separate tracking camera along the barrel, a microcomputer with active cooling, a controller, a compressed gas cylinder, and a distinctive funnel-shaped hopper containing pneumatic bullets.
The key detail is – weaponIt's not a shotgun or a submachine gun, as you might expect given its 20kg total weight. It's pneumatic gun With a high-pressure cylinder and gravity-fed ammunition. Why choose this:
- Low recoil: Lightweight servos stay calibrated after each shot.
- Simple energy regulation: the pressure is changed by a reducer.
- Smooth burst rhythm and predictable ballistics at short ranges.
- The cost of a shot is mere pennies, which is more important for a mass-use system than a beautiful muzzle energy.
The price for this is obvious and is implied in the name itself. For a total weight of 20 kg and a compact cylinder, it's possible to achieve a muzzle energy of around 50-150 joules—the same level as mass-produced 5,5-6,35 mm PCP rifles. .50 PCP systems are lightweight (4-5 kg), but when firing bursts automatically, they're limited not by weight but by two other constraints: air consumption (the cylinder is emptied in tens of shots instead of thousands) and recoil, after which lightweight servos need to be recalibrated. This is unacceptable for a 48-hour standby mode and a "several short bursts at a target" mode. Hence the 25-meter radius: at this distance, the energy is enough to pierce a plastic propeller blade or damage the exposed motor of an FPV drone. The weak points are obvious: the blades (loss of thrust, somersault) and the motors themselves; When hit by a low-energy projectile, the frame and warhead most often simply suffer a dent. Shooting down a drone means hitting the propeller before it reaches the trench. At 25 meters, a 5-inch diameter blade produces an angular displacement of about 0,3°, and keeping the barrel aligned with this target on a moving target is a task not so much for the neural network as for mechanics: drive play, turret inertia, and encoder accuracy are more important here than extra milliseconds of inference. This is primarily a matter of the impact geometry and the quality of the suspension.
Then the neural network takes over. Four overview cameras provide a 360-degree image, a specially trained model identifies the quadcopter's distinctive silhouette in the stream of frames, filters out birds, foliage, and backlight, selects the closest target, and transmits it to the tracking camera. The tracking camera keeps the drone centered in the frame, the servos rotate the turret, and the controller fires a short burst. The operator is not involved in this process.
Figures from the developer's materials: assembled weight is about 20 kg, battery life is up to 48 hours on LiFePO4 batteries, prototype cost is about 230 thousand rubles.
Where the Dome 25 Sag
Next comes what is not in the developer’s press release.
Twenty-five meters is a very narrow window. An FPV drone on final approach flies at a speed of approximately 25-30 m/s, leaving about a second for the system's full response. The components of this budget can be estimated in orders of magnitude. Modern single-board computers like the Jetson Orin, using YOLO family detectors, provide inference of approximately 15-30 milliseconds per frame, depending on the model and input resolution. Added to this are camera latency (frame capture and transmission takes another 20-40 ms), tracker operation, servo commands, and the actual mechanical turret rotation: tens of milliseconds at close range, up to hundreds at wide angles. In total, 100-200 ms for each aiming adjustment is realistic. This is within a second of the drone's flight time, but just barely. The scenario of "the drone emerges from cover five meters away at full speed" is fundamentally impossible, and no neural network can help here.
Optics is always a function of the weather. Fog, heavy rain, dust, smoke over a trench, a low sun directly in the frame—all of these degrade detector performance. A neural network is more robust than traditional video analytics, but it still has to overcome the physics of the sensor. Thermal imaging partially alleviates the situation; it's not yet visible in the publicly available "Kupol 25" layout.
A separate question is why primary detection is entirely reliant on optics. An FPV drone can be heard 100-200 meters before it enters the camera's field of view; a four-channel directional microphone array costs less than a single surveillance camera and provides a rough bearing within tens of milliseconds. This eliminates the main problem—the one-second reaction window—because the turret begins turning to the desired sector even before visually locking on to the target. NeuroPVO itself mentions acoustic detection in its announcements of the future lineup, but it's for aircraft UAVs at ranges of 10-20 km; this channel isn't yet available for the close-in perimeter, where it's most needed. This seems like the system's most obvious redundancy.
The SMASH Hopper is a lightweight, remote-controlled weapon station (turret) manufactured by the Israeli company Smart Shooter.
Similar close-combat turrets with AI recognition are being developed abroad, but there is no direct equivalent to the Kupol 25 among known systems. The Israeli Smash from Smart Shooter is sighting module, which is mounted on a standard assault rifle and helps a human shooter hit a drone; there is no autonomous turret. The German Skynex from Rheinmetall, on the other hand, battery, not turret35mm automatic rifles with programmable AHEAD rounds, truck-mounted deployment, and a unit price in the millions of euros. The closest concept is the American Bullfrog from Allen Control Systems: a robotic turret with automatic drone recognition and a 7,62mm machine gun. The fundamental difference is that the Bullfrog is designed around a combat rifle cartridge. The Kupol 25 uses a pneumatic bullet, and this decision has significant implications: weight, cost, noise, and range. The resulting tactical niches are diverse, yet the underlying philosophy is the same.
Skynex anti-aircraft artillery system produced by the German company Rheinmetall
A separate issue is the ethical and legal one. Press releases typically don't elaborate on this, and forum discussions tend to dismiss it as "defensive." Meanwhile, the UN Group of Governmental Experts on Lethal Autonomous Weapons Systems (GGE on LAWS) has been working within the framework of the Convention on Certain Conventional Weapons since 2017. A unified definition of "meaningful human control" has yet to be developed, but based on established discussion practice, close-combat defensive systems (Phalanx in automatic mode, Israeli Iron Dome, naval AK-630) are usually excluded from discussion. The logic is simple: a human makes decisions at the level of system deployment and operating rules, not each shot fired at a specific target. The Kupol 25 formally falls into the same category: it protects a limited perimeter. However, the combination of "civilian laboratory + unmanned automatic fire + open circuits" is new to international practice, and a regulatory framework for it has not yet been developed.
The main thing is that there is no public data on actual use yet. All that is known are the developer's statements, the layout diagram, and test videos. How many FPV drones the system reliably films in the field, how it performs in rain, how it filters its own reconnaissance copters from other drones—these questions have no publicly available answers.
Where is this heading?
NeuroPVO itself refers to the "Kupol 25" as the first unit of a future line. The "Sbryo 300" system is announced for use against heavy carrier drones at altitudes of up to 300 meters, and work is underway on acoustic detection of fixed-wing UAVs at ranges of 10-20 km. The only change is the target acquisition method: camera, microphone, or radar. After that, it's the same: the model detects, the drive rotates, and the barrel fires.
There is also a second context, without which the whole story It's not readable. The classic answer to FPV is electronic jamming: a jammer disrupts the control channel, the drone loses the operator, and crashes. This system fails where it's physically bypassed: on fiber-optic drones, the channel isn't jammed at all, and on machines with terminal AI guidance, the drone flies the last few hundred meters without communication with the operator, relying on the image from its camera. It's precisely in this niche—where EW powerless—ground-based close-combat turrets become not a supplement, but the only means. The "Dome 25" and its analogs are born not from an abstract love of autonomous weapons, but from a concrete failure of the previous layer of defense.
If by the 1980s, shipborne air defense had developed a layered system from over-the-horizon missiles to Phalanx missiles in the final meters, then ground-based anti-drone air defense is now going through this same process, in fast-forward mode. The idea of a lower level based on AI recognition has already been developed as a class – in Russia, the US, Europe, and Israel simultaneously. The next question is what configuration it will take and who will be the first to bring it to mass production.
The "Kupol 25" concept itself isn't unique; everyone's talking about it right now. The entry barrier is interesting: twenty kilograms, two hundred and thirty thousand rubles for a prototype, and the schematics are publicly available. With such figures, building it in a garage is no longer a figment of the imagination—all you need is a Jetson and someone who knows how to solder. This is where the GGE on LAWS discussion of "autonomous lethal systems" ceases to be a matter of academics: while definitions are being agreed upon in Geneva, a "Kupol 25" equivalent is being assembled in any local radio club. The regulatory framework is noticeably lagging behind engineering reality, and this is perhaps the main long-term theme surrounding such systems—regardless of how effective a particular tower is against a particular FPV.
Information